Uncertainty estimation for models

[mShuntov]

Hi SourceXtractor++ community,

As you probably know by now, I'm extensively using SE++ to make catalogs in multi-band model fitting mode in images from JWST, HST, HSC and UVISTA.

I encounter an issue that becomes more severe in photo-z fitting codes regarding model flux uncertainties in the multi-band fitting mode. In particular, we find that for many faint sources near the depth of the band have a model uncertainty that is very small.
For example, a source that is detected in JWST NIRCam bands (from 1.15 to 4.4 microns) and not in the optical bands can have a U-band flux and error of f=1.9983e-38 ef=8.7781e-52 (mag: 45.6559 mag_err: 4.77056e-14). Such cases can of course be handled and removed 'manually' but that comes at a cost when we just want to compute photo-z for a large catalog.

How are model errors computed for sources? How are models handled in a multi-band fitting case where an object is a dropout in some bands?

Here are some examples of multi-band model fitting with dropouts. As you can see, if a source is detected in the chi2 detection image, a model is forced even in the bands where it is not detected. The scale of the images is automatic, so that's why in the model image it appears very bright, but the residual is usually clean.

It will be very useful if you can give me some insight into how model errors are computed and how this is dealt with in dropout bands.

Thanks!

[mkuemmel]

The model errors come directly from the non-linear minimization algorithm. The error correlations are not taken into account. The photometry plot is very small, but looking at the stamps the error bars on the blue side should be larger.
Our experience is rather positive in the sense that we see model fitting error bars that correlate with image signal. Here an example from the CEERS/JWST data set we work on and which is pretty similar to yours:

Red is the fitted mag and blue is auto_mag (for comparison). Also here the blue side is fainter which shows in error bars. Here some things you could look into:

• if you have pixel-wise RMS/WEIGHT/VARIANCE make sure you specify the correct type;
• if you have no RMS/... check whether RMS derived in SE++ makes sens (checkimage);
• if you have good RMS/... set "gain=0.0" for the measurement images;

Happy to discuss further.

[mShuntov]

Thanks, Martin for answering! Indeed, looking in more detail, on general the model uncertainties correlate with the SNR of the band, modulo individual exceptions that appear to have very small errors but low S/N (whose problem can be traced to other artefacts in the images).

I'll keep you posted on my experience with this. I think this combination of images from different instruments is where SE++ will excel, and it is arguably the only way to do photometric catalogues in the JWST/Euclid/Ground observatories era :)

I have also built catalogues in CEERS combining NIRCam, HST/ACS and WFC and HSC. Would be happy to share a test sample to compare with you so that we improve things :)